6 Myths About GPT-5.6 Sol, Terra, and Luna — and What's Actually True

OpenAI shipped the GPT-5.6 family — Luna, Terra, and Sol — to general availability on July 9, and within hours the hot takes had outrun the facts. Here are six claims making the rounds, checked against the pricing page, the published benchmarks, and early hands-on reports.

Myth 1: "Sol, Terra, and Luna are one model at three price points."

Reality: they are three distinct model sizes, from Luna (smallest) up to Sol (largest). The naming scheme is the actual news here: the version number now identifies the generation, while Sol, Terra, and Luna are durable capability tiers that OpenAI says can advance on their own cadence. Expect a Terra update to ship without waiting for a new Sol.

What they do share: a February 16, 2026 knowledge cutoff, a 1M-token context window, and 128,000 max output tokens across all three.

Myth 2: "The sticker price tells you what you'll pay."

Reality: list pricing per 1M tokens is Luna $1/$6, Terra $2.50/$15, and Sol $5/$30 — but reasoning tokens now dominate real costs. Simon Willison ran the same prompt across all three models at six reasoning-effort levels: the cheapest run (Luna, effort none) cost 0.71 cents, the most expensive (Sol, max) cost 48.55 cents. That is a 68x spread on one prompt.

There is also a premium lane: Sol Fast serves the same flagship weights on Cerebras hardware at up to 750 tokens/second — for $12.50/$75, quadruple the standard Sol rate.

Myth 3: "GPT-5.6 beat Claude across the board."

Reality: it depends entirely on which benchmark you read. On Agents' Last Exam, a long-horizon agentic eval spanning 55 professional fields, Sol posts a new high of 53.6 — 13.1 points ahead of Claude Fable 5. But on SWE-Bench Pro, OpenAI's own reported numbers have Fable 5 at 80% versus Sol's 64.6%. And on Terminal-Bench 2.1, base Sol's 88.8% edges Claude Mythos 5's 88.0% by less than a point.

The honest summary: GPT-5.6 leads on long-running agentic work and cost efficiency; Anthropic still leads on the hardest coding benchmarks. Nobody swept.

Myth 4: "OpenAI debunked SWE-Bench Pro anyway."

Reality: not yet. The day before launch, OpenAI published an audit estimating that ~30% of SWE-Bench Pro tasks are broken and advising developers to examine results carefully. The audit may well be right — benchmark rot is real — but publishing it 24 hours before shipping a model that trails on that exact benchmark earns a grain of salt. Treat it as a claim awaiting independent replication, not a settled verdict.

Myth 5: "It's just a weights bump — the API didn't change."

Reality: the API changes may matter more than the benchmark deltas. Four additions worth knowing:

Programmatic Tool Calling — the model composes and runs JavaScript that orchestrates tool calls, instead of round-tripping each call through you.
Multi-agent — spinning up subagents for parallel focused work is now a first-class API primitive, not a framework pattern.
Prompt cache breakpoints — explicit, Claude-style cache markers alongside the existing automatic detection.
`detail: original` — image inputs can now skip resizing entirely.

Model strings follow the tier names:

bash

curl https://api.openai.com/v1/responses \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -d '{"model": "gpt-5.6-terra", "input": "..."}'

Myth 6: "Luna is the toy tier."

Reality: Luna gets the same 1M context window, the same 128K output ceiling, and the same knowledge cutoff as Sol. OpenAI claims Luna and Terra both outperform Claude Fable 5 on Agents' Last Exam at roughly one-sixteenth the cost — self-reported, but consistent with the family's efficiency framing. Terra, meanwhile, matches GPT-5.5-level performance at half the price.

The practical playbook: start on Terra, drop to Luna where quality holds, and reserve Sol — with reasoning effort capped — for the work that actually needs it.